What Is an Archaeon and Are the Archaea Really Unique?

1 What is an archaeon and are the

2 Archaea really unique?

1* 3 Ajith Harish Department of Cell and Molecular Biology, Structural and Molecular Biology Program, *For correspondence: 4 1 [email protected] Uppsala University, Uppsala, Sweden 5

7 Abstract The recognition of the group Archaea 40 years ago stimulated research in microbial 8 evolution and molecular systematics that prompted a new classificatory scheme to organize 9 biodiversity. Advances in DNA sequencing techniques have since significantly improved the 10 genomic representation of the archaeal biodiversity. In addition, advances in phylogenetic 11 modeling that facilitate large-scale phylogenomics have resolved many recalcitrant branches of the 12 Tree of Life. Despite the technical advances and an expanded taxonomic representation, two 13 important aspects of the origins and evolution of the Archaea remain controversial, even as we 14 celebrate the 40th anniversary of the monumental discovery. The issues concern (i) the uniqueness 15 (monophyly) of the Archaea, and (ii) the evolutionary relationships of the Archaea to the Bacteria 16 and the Eukarya; both of these are relevant to the deep structure of the Tree of Life. The 17 uncertainty is primarily due to a scarcity of information in standard datasets—the core-genes 18 datasets—to reliably resolve the conflicts. These conflicts can be resolved efficiently by employing 19 complex genomic features and genome-scale evolution models—a distinct class of phylogenomic 20 characters and evolution models—that can be employed routinely to maximize the use of genome 21 sequences as well as to minimize uncertainties in tests of evolutionary hypotheses.

23 Introduction 24 The recognition of the Archaea as the so-called “third form of life” was made possible in part by a 25 new technology for sequence analysis, oligonucleotide cataloging, developed by Fredrik Sanger and 26 colleagues in the 1960s (1, 2). Carl Woese’s insight of using this method, and the choice of the small 27 subunit ribosomal RNA (16S/SSU rRNA) as a phylogenetic marker, not only put microorganisms 28 on a phylogenetic map (or tree), but also revolutionized the field of molecular systematics that 29 Zukerkandl and Pauling has previously alluded to (3). Comparative analysis of organism-specific 30 (oligonucleotide) sequence-signatures in SSU rRNA led to the recognition of a distinct group of 31 microorganisms (2, 4). Initially referred to as Archaeabacteria, these unusual organisms had 32 ‘oligonucleotide signatures’ distinct from other bacteria (Eubacteria), and they were later found to 33 be different from those of Eukarya (eukaryotes) as well. Many other features, including molecular, 34 biochemical as well as ecological, corroborated the uniqueness of the Archaea. Thus the archaeal 35 concept was established (2). 36 The study of microbial diversity and evolution has come a long way since then: sequencing 37 microbial genomes, and directly from the environment without the need for culturing is now 38 routine (5, 6). This wealth of sequence information is exciting not only for cataloging and organizing 39 biodiversity, but also to understand the ecology and evolution of microorganisms – archaea and 40 bacteria as well as eukaryotes – that make up a vast majority of the planetary biodiversity. Since 41 large-scale exploration by the means of environmental genome sequencing became possible almost 42 a decade ago, there has also been a palpable excitement and anticipation of the discovery of a

1 of 21 43 fourth form of life or a “fourth domain” of life (7). The reference here is to a fourth form of cellular 44 life, but not to viruses, which some have already proposed to be the fourth domain of the Tree of 45 Life (ToL) (7, 8). If a fourth form of life were to be found, what would the distinguishing features be, 46 and how could it be measured, defined and classified? 47 Rather than the discovery of a fourth domain, and contrary to the expectations, however, current 48 discussion is centered around the return to a dichotomous classification of life (9-11), despite the 49 rapid expansion of sequenced biodiversity – hundreds of novel phyla descriptions (12, 13). The 50 proposed dichotomous classifications schemes, unfortunately, are in sharp contrast to each other, 51 depending on: (i) whether the Archaea constitute a monophyletic group—a unique line of descent 52 that is distinct from those of the Bacteria as well as the Eukarya; and (ii) whether the Archaea form 53 a sister clade to the Eukarya or to the Bacteria. Both the issues stem from difficulties involved in 54 resolving the deep branches of the ToL (10, 11, 14). 55 The twin issues, first recognized in the 80s based on single-gene (SSU rRNA) analyses, continue 56 to be the subjects of a long-standing debate, which remains unresolved despite large-scale analyses 57 of multi-gene datasets (5, 15-19). In addition to the choice of genes to be analyzed, the choice of 58 the underlying character evolution model is at the core of contradictory results that either supports 59 the Three-domains tree (5, 19) or the Eocyte tree (17, 20). In many cases, adding more data, either 60 as enhanced taxon (species) sampling or enhanced character (gene) sampling, or both, can resolve 61 ambiguities (21, 22). However, as the taxonomic diversity and evolutionary distance increases 62 among the taxa studied, the number of conserved marker-genes that can be used for phylogenomic 63 analyses decreases. Accordingly, resolving the phylogenetic relationships of the Archaea, Bacteria 64 and Eukarya is restricted to a small set of genes—50 at most—in spite of the large increase in the 65 numbers of genomes sequenced and the associated development of sophisticated phylogenomic 66 methods. 67 Based on a closer scrutiny of the recent phylogenomic datasets employed in the ongoing 68 debate, I will show here that one of the reasons for this persistent ambiguity is that the ‘information’ 69 necessary to resolve these conflicts is practically nonexistent in the standard marker-genes (i.e. core- 70 genes) datasets employed routinely for phylogenomics. Further, I discuss analytical approaches 71 that maximize the use of the information that is in genome sequence data and simultaneously 72 minimize phylogenetic uncertainties. In addition, I discuss simple but important, yet undervalued, 73 aspects of phylogenetic hypothesis testing, which together with the new approaches hold promise 74 to resolve these long-standing issues effectively.

75 Results 76 Information in core genes is inadequate to resolve the archaeal radiation 77 Data-display networks (DDNs) are useful to examine and visualize character conflicts in phylogenetic 78 datasets, especially in the absence of prior knowledge about the source of such conflicts, ideally 79 before downstream processing of the data for phylogenetic inference (23, 24). While congruent 80 data will be displayed as a tree in a DDN, incongruences are displayed as reticulations in the tree. 81 Fig. 1A shows a neighbor-net analysis of the SSU rRNA alignment used to resolve the phylogenetic 82 position of the recently discovered Asgard archaea (20). The DDN is based on character distances 83 calculated as the observed genetic distance (p-distance) of 1,462 characters, and shows the total 84 amount of conflict in the dataset that is incongruent with character bipartitions (splits). The edge 85 (branch) lengths in the DDN correspond to the support for the respective splits. Accordingly, two 86 well-supported sets of splits for the Bacteria and the Eukarya are observed. The Archaea, however, 87 does not form a distinct, well-resolved/well-supported group, and is unlikely to correspond to a 88 monophyletic group in a phylogenetic tree.

2 of 21 Figure 1. Data-display networks depicting the character conflicts in different datasets that employ different character types. (A) SSU rRNA alignment of 1,462 characters. Concatenated protein sequence alignment of (B) 29 core-genes, 8,563 characters; (C) 48 core-genes, 9,868 characters and (D) also 48 core-genes, 9,868 SR4 recoded characters (data simplified from 20 to 4 character-states). Each network is constructed from a neighbor-net analysis based on the observed genetic distance (p-distance) and displayed as an equal angle split network. Edge (branch) lengths correspond to the support for character bipartitions (splits), and reticulations in the tree correspond to character conflicts. Datasets in (A), (C) and (D) are from Ref. 20, and in (B) is from Ref. 17.

89 Likewise, the concatenated protein sequence alignment of the so-called ‘genealogy deﬁning core 90 of genes’ (25) – a set of conserved single-copy genes – also does not support a unique archael lineage. 91 Fig. 1B is a DDN derived from a neighbor-net analysis of 8,563 characters in 29 concatenated core- 92 genes (17), while Fig. 1C,D is based on 9,868 characters in 44 concatenated core-genes (also from 93 (20)). Even taken together, none of the standard marker gene datasets are likely to support the 94 monophyly of the Archaea — a key assertion of the three-domains hypothesis (26). Simply put, 95 there is not enough information in the core-gene datasets to resolve the archaeal radiation, or to

3 of 21 96 determine whether the Archaea are really unique compared to the Bacteria and Eukarya. However, 97 other complex features — including molecular, biochemical and phenotypic characters, as well 98 as ecological adaptations — support the uniqueness of the Archaea. These idiosyncratic archaeal 99 characters include the subunit composition of supramolecular complexes like the ribosome, DNA- 100 and RNA-polymerases, biochemical composition of cell membranes, cell walls, and physiological 101 adaptations to energy-starved environments, among other things (27, 28).

102 Complex phylogenomic characters minimize uncertainties regarding the unique- 103 ness of the Archaea 104 A nucleotide is the smallest possible locus, and an amino acid is a proxy for a locus of a nucleotide 105 triplet. Unlike the elementary amino acid- or nucleotide-characters in the core-genes dataset (Fig.1), 106 the DDN in Fig. 2 is based on complex molecular characters – genomic loci that correspond to 107 protein domains, typically ~200 amino acids (600 nucleotides) long. Neighbor-net analysis of protein- 108 domain data coded as binary characters (presence/absence) is based on the Hamming distance 109 (identical to the p-distance used in Fig.1). Here the Archaea also form a distinct well-supported 110 cluster, as do the Bacteria and the Eukarya.

Figure 2. Data-display networks (DDN) depicting character conﬂicts among complex phylogenomic characters – genomic loci corresponding to protein-domains in this case. (A) Neighbor-net analysis based on Hamming distance (identical to the p-distance used in Fig.1) of 1,732 characters sampled from 141 species. (B) DDN based on an enriched taxon sampling of 81 additional species totaling 222 species and a modest increase to 1,738 characters. The dataset in (A) is from Ref. 10, which was updated with novel species to represent the recently described archaeal and bacterial species (5, 12, 20).

111 Fig 2A is a DDN based on the dataset that includes protein-domain cohorts of 141 species, used 112 in a phylogenomic analysis to resolve the uncertainties at the root of the ToL (29). Compared to 113 the data in Fig. 1, the taxonomic diversity sampled for the Bacteria and Eukarya is more extensive, 114 but less extensive for the Archaea; it is composed of the traditional groups Euryarchaeota and 115 Crenarchaeota.Fig. 2B is a DDN of an enriched sampling of 81 additional species, which includes 116 representatives of the newly described archaeal groups: TACK (30), DPANN (5), and Asgard group

4 of 21 117 including the Lokiarchaeota (20). In addition, species sampling was enhanced with representatives 118 from the candidate phyla described for Bacteria, and with unicellular species of Eukarya. The 119 complete list species analyzed is in SI Table 1. 120 Notably, the extension of the protein-domain cohort was insigniﬁcant, from 1,732 to 1,738 121 distinct domains (characters). Based on the well-supported splits in the DDN that form a distinct 122 archaeal cluster, the Archaea are likely to be a monophyletic group (clade) in phylogenies inferred 123 from these datasets.

124 Data quality affects model complexity required to explain phylogenetic datasets 125 Resolving the paraphyly or monophyly of the Archaea is relevant to determining whether the Eocyte 126 tree (Fig. 3A) or the Three-domains tree (Fig. 3B), respectively, is a better-supported hypothesis. 127 Recovering the Eocyte tree typically requires implementing complex models of sequence evolution 128 rather than their relatively simpler versions (11). In general, complex models tend to fit the data 129 better. For instance, according to a model selection test for the 29 core-genes dataset, the LG 130 model (31) of protein sequence evolution is a better-fitting model than other standard models, 131 such as the WAG or JTT substitution model (SI-Table 2), as reported previously (17). Further, a 132 relatively more complex version of the LG model, with multiple rate-categories was found to be a 133 better-fitting model than the simpler single-rate-category model (Fig. 3C; SI-Table 2). The fit of the 134 data is estimated as the likelihood of the best tree given the model. 135 A complex, multiple rate-categories model accounts for site-specific substitution rate variation. 136 Substitution-rate heterogeneity across different sites in the multiple-sequence alignment (MSA) 137 was approximated using a discrete Gamma model with 4, 8 or 12 rate categories (LG+G4, LG+G8 or 138 LG+G12, respectively). The Archaea is consistent with a paraphyletic group in trees derived from the 139 rate-heterogeneous versions of the LG model (Fig. 3A). Furthermore, the fit of the data improves 140 with the increase in complexity of the substitution model (Fig. 3C). Model complexity increases 141 with any increase in the number of rate categories and/or the associated numbers of parameters 142 that need to be estimated. However, with a relatively simpler version – a rate-homogeneous LG 143 model, in which the substitution-rates are approximated to a single rate-category, the Archaea are 144 consistent with a monophyletic group (Fig. 3B). 145 In contrast, trees inferred from the protein-domain datasets are consistent with monophyly 146 of the Archaea irrespective of the complexity of the underlying model (Fig. 3D-F). The Mk model 147 (Markov k model) is the best-known probabilistic model of discrete character evolution, particularly 148 of complex characters coded as binary-state characters (32, 33). Since the Mk model assumes a 149 stochastic process of evolution, it is able to estimate multiple state changes along the same branch. 150 Implementing a simpler rate-homogeneous version of the Mk model (Fig. 3D), as well as more 151 complex rate-heterogeneous versions with 4, 8 or 12 rate categories (Mk+G4, Mk+G8 or Mk+G12, 152 respectively), also recovered trees that are consistent with the monophyly of the Archaea (Fig. 3E) 153 The tree derived from the Mk+G4 model is shown in Fig. 3E. While the tree derived from Mk+G8 154 model is identical (SI-Fig. 1) to the Mk+G4 tree, the Mk+G12 tree is almost identical with minor 155 differences in the bacterial sub-groups (SI-Fig. 2) 156 In all cases, bipartitions for Archaea show strong support with posterior probability (PP) of 0.99 157 while that of Bacteria and Eukarya is supported with a PP of 1.0; in spite of substantially different 158 fits of the data. The uniqueness of the Archaea is almost unambiguous in this case (but see next 159 section).

5 of 21 Figure 3. Comparison of concatenated-gene trees derived from amino acid characters and genome trees derived from protein-domain characters. Branch support is shown only for the major branches. Scale bars represent the expected number of changes per character. (A), (B) Core-genes-tree derived from a better-fitting model (LG+G4) and a worse fitting mode (LG), respectively, of amino acid substitutions. (C) Model fit to data is ranked according the log likelihood ratio (LLR) scores. LLR scores are computed as the difference from the best-fitting model (LG+G12) of the likelihood scores estimated in PhyML. Thus, larger LLR values indicate less support for that model/tree relative to the most-likely model/tree. Substitution rate heterogeneity is approximated with 4, 8 or 12 rate categories in the complex models, but with a single rate category in the simpler model. (D), (E) are genome-trees derived from a better-fitting model (Mk+G4) and a worse fitting model (Mk), respectively, of protein-domain innovation. (F) Model fit to data is ranked according log Bayes factor (LBF) scores, which like LLR scores are the log odds of the hypotheses. LBF scores are computed as the difference in likelihood scores estimated in MrBayes.

6 of 21 160 Siblings and cousins are indistinguishable when reversible models are employed 161 Although a DDN is useful to identify and diagnose character conﬂicts in phylogenetic datasets and 162 to postulate evolutionary hypotheses, a DDN by itself cannot be interpreted as an evolutionary 163 network, because the edges do not necessarily represent evolutionary phenomena and the nodes 164 do not represent ancestors (23, 24). Therefore, evolutionary relationships cannot be inferred from 165 a DDN. Likewise, evolutionary relationships cannot be inferred from unrooted trees, even though 166 nodes in an unrooted tree do represent ancestors and an evolution model deﬁnes the branches 167 (see Fig. 4A).

Figure 4. Effect of alternative ad hoc rootings on the phylogenetic classification of archaeal biodiversity. (A) An unrooted tree is not fully resolved into bipartitions at the root of the tree (i.e. a polytomous rather than a dichotomous root branching) and thus precludes identification of sister group relationships. It is common practice to add a user-specified root a posteriori based on prior knowledge (or belief) of the investigator. Four possible (of many) rootings R1-R4 are shown. (B) Operationally, adding a root (rooting) a posteriori amounts to adding new information – a new bipartition and an ancestor as well as an evolutionary polarity – that is independent of the source data. (C-F) The different possible evolutionary relationships of the Archaea to other taxa, depending on the position of the root, are shown. Rooting is necessary to determine the recency of common ancestry as well the temporal order of key evolutionary transitions that define phylogenetic relationships.

168 An unrooted tree, unlike a rooted tree, is not an evolutionary (phylogenetic) tree per se, since it 169 is a minimally deﬁned hypothesis of evolution or of relationships; it is, nevertheless, useful to rule

7 of 21 170 out many possible bipartitions and groups (34, 35). Given that a primary objective of phylogenetic 171 analyses is to identify clades and the relationships between these clades, it is not possible to 172 interpret an unrooted tree meaningfully without rooting the tree (see Fig. 4A). Identifying the root is 173 essential to: (i) distinguish between ancestral and derived states of characters, (ii) determine the 174 ancestor-descendant polarity of taxa, and (iii) diagnose clades and sister-group relationships (Fig. 175 4). Yet, most phylogenetic software construct only unrooted trees, which are then consistent with 176 several rooted trees (Fig. 4 C-F). However, an unrooted tree cannot be fully resolved into bipartitions, 177 because an unresolved polytomy (a trifurcation in this case) exists near the root of the tree (Fig. 4A), 178 which otherwise corresponds to the deepest split (root) in a rooted tree (Fig. 4, C-F). 179 Resolving the polytomy requires identifying the root of the tree. The identity of the root 180 corresponds, in principle, to any one of the possible ancestors as follows: 181 i. Any one of the inferred-ancestors at the resolved bipartitions (open circles in Fig. 4A), or 182 ii. Any one of the yet-to-be-inferred-ancestors that lies along the stem-branches of the unre- 183 solved polytomy (dashed lines in Fig. 4A) or along the internal-braches. 184 In the latter case, rooting the tree a posteriori on any of the branches amounts to inserting an ad- 185 ditional bipartition and an ancestor that is neither inferred from the source data nor deduced from 186 the underlying character evolution model. Since standard evolution models employed routinely 187 cannot resolve the polytomy, rooting, and hence interpreting the Tree of Life depends on: 188 i. Prior knowledge — eg., fossils or a known sister-group (outgroup), or 189 ii. Prior beliefs/expectations of the investigators — eg., simple is primitive (36, 37), bacteria are 190 primitive (38, 39), archaea are primitive (1), etc. 191 Both of these options are independent of the data used to infer the unrooted ToL. Some possible 192 rootings and the resulting rooted-tree topologies are shown as cladograms in Fig. 4, C-F. If the root 193 lies on any of the internal branches (e.g. R1 in Fig. 4,A-C), or corresponds to one of the internal 194 nodes, within the archael radiation, the Archaea would not constitute a unique clade (Fig. 4C). 195 However, if the root lies on one of the stem-branches (R2/R3/R4 in Fig. 4 A, B), monophyly of 196 the Archaea would be unambiguous (Fig. 4 D-F). Determining the evolutionary relationship of the 197 Archaea to other taxa, though, requires identifying the root. 198 Directional evolution models, unlike reversible models, are able to identify the polarity of state 199 transitions, and thus the root of a tree (40-42). Therefore, the uncertainty due to a polytomous root 200 branching is not an issue (Fig 5A). Moreover, directional evolution models are useful to evaluate the 201 empirical support for prior beliefs about the universal common ancestor (UCA) at the root of the 202 ToL (29). A Bayesian model selection test implemented to detect directional trends (42) chooses the 203 directional model, overwhelmingly (Fig. 5B), over the unpolarized model for the protein-domain 204 dataset in Fig. 2B, as reported previously for the dataset in Fig. 2A (29). Further, the best-supported 205 rooting corresponds to root R4 (Fig. 4F and Fig. 5A) — monophyly of the Archaea is maximally 206 supported (PP of 1.0). Furthermore, the sister-group relationship of the Archaea to the Bacteria 207 is maximally supported (PP 1.0). Accordingly, a higher order taxon, Akaryotes, proposed earlier 208 (Forterre 1992) forms a well-supported clade. Thus Akaryotes (or Akarya) and Eukarya are sister 209 clades that diverge from the UCA at the root of the ToL, also as reported previously (29). 210 Alternative rootings are much less likely, and are not supported (Fig. 5C). Accordingly, indepen- 211 dent origin of the eukaryotes as well akaryotes is the best-supported scenario. The Three-domains 171 212 tree (root R3, Fig. 4E) is 10 times less likely, and the scenario proposed by the Eocyte hypothesis 213 (root R1, Fig. 5A) is highly unlikely. The common belief that simple is primitive, as well as beliefs that 214 archaea are primitive or that archaea and bacteria evolved before eukaryotes, are not supported 215 either.

216 Employing complex molecular characters maximizes representation of orthologous, 217 non-recombining genomic loci, and thus phylogenetic signal 218 Genomic loci that can be aligned with high conﬁdence using MSA algorithms are typically more 219 conserved than those loci for which alignment uncertainty is high. Such ambiguously aligned

8 of 21 Figure 5. (A) Rooted tree of life inferred from patterns of inheritance of unique genomic-signatures. A dichotomous classification of the diversity of life such that Archaea is a sister group to Bacteria, which together constitute a clade of akaryotes (Akarya). Eukarya and Akarya are sister-clades that diverge from the root of the tree of life. Each clade is supported by the highest posterior probability of 1.0. The phylogeny supports a scenario of independent origins and descent of eukaryotes and akaryotes. (B) Model selection tests identify, overwhelmingly, directional evolution models to be better-fitting models. (C) Alternative rootings, and accordingly alternative classifications or scenarios for the origins of the major clades of life, are much less probable and not supported.

9 of 21 220 regions of sequences are routinely trimmed off before phylogenetic analyses (43). Typically, 221 the conserved well-aligned regions correspond to protein domains with highly ordered three- 222 dimensional (3D) structures with specific 3D folds (Fig. 6A). Regions of sequences that are trimmed 223 usually show higher variability in length, are less ordered and are known to accumulate insertion 224 and deletion (indel) mutations at a higher frequency than in the regions that correspond to folded 225 domains (44). These variable, structurally disordered regions, which flank the structurally ordered 226 domains, link different domains in multi-domain proteins (Fig. 6A). Multi-domain architecture (MDA), 227 the N-to-C terminal sequence of domain arrangement, is distinct for a protein family, and differs in 228 closely related protein families with similar functions (Fig. 6A). The variation in MDA also relates to 229 alignment uncertainties.

Figure 6. Alignment uncertainty in closely related proteins due to domain recombination. (A) Multi-domain architecture (MDA) of the translational GTPase superfamily based on recombination of 8 modular domains. 57 distinct families with varying MDAs are known, of which 6 canonical families are shown as a schematic on the left and the corresponding 3D folds on the right. Amino acid sequences of only 2 of the 8 conserved domains can be aligned with conﬁdence for use in phylogenetic analysis. The length of the alignment varies from 200-300 amino acids depending on the sequence diversity sampled (14,76). The EF-Tu—EF-G paralogous pair employed as pseudo-outgroups for the classical rooting of the rRNA tree is highlighted. (B) Phyletic distribution of 1,738 out the 2,000 distinct SCOP-domains sampled from 222 species used for phylogenetic analyses in the present study. About 70 percent of the domains are widely distributed across the sampled taxonomic diversity.

10 of 21 230 A closer look at the 29 core-genes dataset shows that the concatenated-MSA corresponds to a 231 total of 27 distinct protein domains or genomic loci (Table 1). The number of loci sampled from 232 diﬀerent species varies between 20 and 27, since not all loci are found in all species. While some 233 loci are absent in some species, some loci are redundant. For instance, the P-loop NTP hydrolase 234 domain, one of the most prevalent protein domains, is represented up to 9 times in many species 235 (Table 1). Many central cellular functions are driven by the conformational changes in proteins 236 induced by the hydrolysis of nucleoside triphosphate (NTP) catalyzed by the P-loop domain. Out of 237 a total of 27 distinct domains, 7 are redundant, with two or more copies represented per species. 238 Similarly, 9 of the 50 domains have a redundant representation in the 44 core-gene dataset (Table 1). 239 The observed redundancy of the genomic loci in the core-genes alignments is inconsistent with the 240 common (and typically untested) assumption of using single-copy genes as a proxy for orthologous 241 loci sampled for phylogenetic analysis.

No. of Redundant domains No. of times No. of taxa unique No. of redundant in in which Dataset No. of taxa genes unique domains each taxon redundant SCOP Unique ID Description 9 29 52540 P-loop containing NTP hydrolases 8 6 7 2 3 10 29 core-genes 50447 Translation proteins 44 29 27 2 13 dataset 3 33 54211 Ribosomal protein S5 domain 2-like 2 4 50249 Nucleic acid-binding proteins 2 17 EF-G C-terminal domain-like 2 34 64484 beta and beta-prime subunits of DNA dependent RNA-polymerase 2 37 5 3 4 81 50249 Nucleic acid-binding proteins 3 11 2 1 3 78 50104 Translation proteins SH3-like domain 2 18 3 15 48 core-genes 50447 Translation proteins 96 48 50 2 71 dataset 3 88 64484 beta and beta-prime subunits of DNA dependent RNA-polymerase 2 5 52540 P-loop containing NTP hydrolases 2 83 53067 Actin-like ATPase domain 2 40 53137 Translational machinery components 2 90 54211 Ribosomal protein S5 domain 2-like 2 93 56053 Ribosomal protein L6 2 88

Figure 7. Redundant representation of protein-domains in concatenated core-genes datasets. The P-loop NTP hydrolase domain is one of the most prevalent domain. Genomic loci corresponding to P-loop hydrolase domain are represented 8-9 times in each species in the single-copy genes employed from core-genes multiple sequence alignments. Redundant loci in the core-genes datasets vary depending on the genes and species sampled for phylogenomic analyses.

242 In contrast, the protein-domain datasets are composed of unique loci (Fig. 6B). Despite the 243 superficial similarity of the DDNs in Fig.1 and Fig.2, they are both qualitatively and quantitatively 244 different codings of genome sequences. As opposed to tracing the history of 30-50 loci in the 245 standard core-genes datasets (Fig. 1), up to 60 fold (1738 loci) more information can be represented 246 when genome sequences are coded as protein-domain characters (Fig. 2). Currently 2,000 unique 247 domains are described by SCOP (Structural Classification of Proteins) (45). The phyletic distribution 248 of 1,738 domains identified in the 222 representative species sampled here is shown in a Venn 249 diagram (Fig. 5B).

250 Discussion 251 Improving data quality can be more eﬀective for resolving recalcitrant branches 252 than increasing model complexity 253 In the phylogenetic literature, the concept of data quality refers to the quality or the strength of 254 the phylogenetic signal that can be extracted from the data. The strength of the phylogenetic 255 signal is proportional to the conﬁdence with which unique state-transitions can be determined for 256 a given set of characters on a given tree. Ideally, historically unique character transitions that entail

11 of 21 257 rare evolutionary innovations are desirable, to identify patterns of uniquely shared innovations 258 (synapomorphies) among lineages. Synapomorphies are the diagnostic features used for assessing 259 lineage-specific inheritance of evolutionary innovations. Therefore identifying character transitions 260 that are likely to be low probability events is a basic requirement for the accuracy of phylogenetic 261 analysis. 262 In their pioneering studies, Woese and colleagues identified unique features of the SSU rRNA 263 – [oligonucleotide] “signatures” – that were six nucleotides or longer, to determine evolutionary 264 relationships (2). An underlying assumption was that the probability of occurrence of the same set 265 of oligomer signatures by chance, in non-homologous sequences, is low in a large molecule like 266 SSU rRNA (1500-2000 nucleotides). Oligomers shorter than six nucleotides were statistically less 267 likely to be efficient markers of homology (46). Thus SSU rRNA was an information-rich molecule to 268 identify homologous signatures (characters) useful for phylogenetic analysis. 269 However, as sequencing of full-length rRNAs and statistical models of nucleotide substitution 270 became common, complex oligomer-characters were replaced by elementary nucleotide-characters; 271 and more recently by amino acid characters. Identifying rare or historically unique substitutions in 272 empirical datasets has proven to be difficult (47, 48), consequently the uncertainty of resolving the 273 deeper branches of the Tree of Life using marker-gene sequences remains high. A primary reason 274 is the prevalence of phylogenetic noise (homoplasy) in primary sequence datasets (Figs 1), due to 275 the characteristic redundancy of nucleotide and amino acid substitutions and the resulting difficulty 276 in distinguishing phylogenetic noise from signal (homology) (49, 50). Better-fitting (or best-fitting) 277 models are expected to extract phylogenetic signal more efficiently and thus explain the data better, 278 but tend to be more complex than worse-fitting models (Fig. 3 C, F). Increasingly sophisticated 279 statistical models that have been developed over the years have only marginally improved the 280 situation (51, 52). Although increasing model complexity can correct errors of estimation and 281 improve the fit of the data to the tree, it is not a solution to improve phylogenetic signal, especially 282 when not present in the source data. 283 Character recoding is found to be effective in reducing the noise/redundancy in the data, and 284 thus uncertainties in phylogenetic reconstructions. This is a form of data simplification wherein 285 the number of amino acid alphabets is reduced to a smaller set of alphabets that are frequently 286 substituted for each other, usually reduced from 20 to 6. Character recoding into reduced alphabets 287 is useful in cases were compositional heterogeneity or substitution saturation is high. However, 288 datasets in which phylogenetic noise is inherently limited are more desirable, to minimize ambi- 289 guities. Like amino acids, protein domains are also modular alphabets, albeit higher order and 290 more complex alphabets of proteins. Moreover, unlike the 20 standard amino acids, there are 291 approximately 2,000 unique protein domains identified at present according to SCOP (45). The 292 number is expected to increase; the theoretical estimates range between 4,000 and 10,000 distinct 293 domain modules, depending on the classification scheme (53). Coding features as binary characters 294 is the simplest possible representation of data for describing historically unique events. 295 The idea of ‘oligonucleotide-signatures’ used for estimating a gene phylogeny has been extended, 296 naturally, to infer a genome phylogeny (54). The signatures were defined in terms of protein-coding 297 genes that were shared among the Archaea. However, as proteins are mosaics of domains, domains 298 are unique genomic signatures (Fig. 6). Protein domains defined by SCOP correspond to complex 299 ‘multi-dimensional signatures’ defined by: (i) a unique 3D fold, (ii) a distinct sequence profile, and 300 (iii) a characteristic function. Though domain recombination is frequent, substitution of one protein 301 domain for another has not been observed in homologous proteins (Fig. 6). For phylogenomic 302 applications protein domains are ‘sequence signatures’ that essentially correspond to single-copy 303 orthologous loci when coded as binary-state characters (presence/absence). These sequence 304 signatures are consistent with unique, non-recombining genomic loci, and are identified using 305 sophisticated statistical models — profile hidden Markov models (pHMMs) (55, 56) — that can be 306 used routinely to annotate and curate genome sequences in automated pipelines (57, 58). 307 For these reasons, protein domains are ideal molecular phylogenetic markers for which character-

12 of 21 308 homology can be validated through more than one property, statistically signiﬁcant (i) sequence 309 similarity, (ii) 3D structure similarity; and (iii) function similarity. In addition, employing genomic loci 310 for protein domains maximizes the genomic information that can be employed for phylogenetic 311 analysis. Even though many other genomic features are known to be useful markers (59), protein 312 domains are the most conserved as well as most widely applicable genomic characters (Fig. 6B).

313 Sorting vertical evolution (signal) and horizontal evolution (noise) 314 Single-copy genes are employed as phylogenetic markers to minimize phylogenetic noise caused 315 by reticulate evolution, including hybridization, introgression, recombination, horizontal transfer 316 (HT), duplication-loss (DL), or incomplete lineage sorting (ILS) of genomic loci. However, the noise 317 observed in the DDNs based on MSA of core-genes (Fig. 1) cannot be directly related to any of the 318 above genome-scale reticulations, since the characters are individual nucleotides or amino acids. 319 Apart from stochastic character conflicts, the observed conflicts are better explained by convergent 320 substitutions, given the redundancy of substitutions. Convergent substitutions caused either due 321 to stringent selection or by chance are a well-recognized form of homoplasy in gene-sequence data 322 (47, 50, 60), and based on recent genome-scale analyses it is now known to be rampant (61, 62). 323 The observed noise in the DDNs based on protein-domain characters (Fig. 2), however, can be 324 related directly to genome-scale reticulation processes and homoplasies. In general, homoplasy 325 implies evolutionary convergence, parallelism or character reversals caused by multiple processes. 326 In contrast, homology implies only one process: inheritance of traits that evolved in the common 327 ancestor and were passed to its descendants. Operationally, tree-based assessment of homol- 328 ogy requires tracing the phylogenetic continuity of characters (and states), whereas homoplasy 329 manifests as discontinuities along the tree. Since clades are diagnosed on the basis of shared 330 innovations (synapomorphies) and defined by ancestry (63, 64), accuracy of a phylogeny depends 331 on an accurate assessment of homology — unambiguous identification of relative synapomorphies 332 on a best fitting tree. 333 Identifying homoplasies caused by character reversals, i.e. reversal to ancestral states requires 334 identification of the ancestral state of the characters under study. However, implementing reversible 335 models precludes the estimation of ancestral states, in the absence of sister groups (outgroups) 336 or other external references. Thus, the critical distinction between shared ancestral homology 337 (symplesiomorphy) and shared derived homology (synapomorphy) is not possible with unrooted 338 trees derived from standard reversible models. Hence, unrooted trees (Fig. 3) are not evolutionary 339 (phylogenetic) trees per se, as they are uninformative about the evolutionary polarity (34, 35, 65). 340 Thus, identifying the root (or root-state) is crucial to (i) determine the polarity of state transitions, (ii) 341 identify synapomorphies, and (iii) diagnose clades. 342 Moreover, because clades are associated with the emergence and inheritance of evolutionary 343 novelties, the discovery of clades is fundamental for describing and diagnosing sister group dif- 344 ferences, which is a primary objective of modern systematics (66). A well-recognized deficiency of 345 phylogenetic inference based on primary sequences is the abstraction of evolutionary ‘information’ 346 (54), often into less tangible quantitative measures. For instance, ‘information’ relevant to diag- 347 nosing clades and support for clades is abstracted to branch lengths. Branch-length estimation is, 348 ideally, a function of the source data and the underlying model. However, in the core-genes dataset 349 the estimated branch lengths and the resulting tree is an expression of the model rather than of the 350 data (Fig. 3 A, B). Some pertinent questions then are: should diagnosis of clades and the features 351 by which clades are identified be delegated to, and restricted to, substitution mutations in a small 352 set of loci and substitution models? Are substitution mutations in 40-50 loci more informative, or 353 the birth and death of unique genomic loci more informative? 354 Proponents of the total evidence approach recommend that all relevant information — molecu- 355 lar, biochemical, anatomical, morphological, fossils — should be used to reconstruct evolutionary 356 history, yet genome sequences are the most widely applicable data at present (59, 67). Accordingly, 357 phylogenetic classification is, in practice, a classification of genomes. There is no a priori theoretical

13 of 21 358 reason that phylogenetic inference should be restricted to a small set of genomic loci corresponding 359 to the core genes, nor is there a reason for limiting phylogenetic models to interpreting patterns 360 of substitution mutations alone. The ease of sequencing and the practical convenience of assem- 361 bling large character matrices, by themselves, are no longer compelling reasons to adhere to the 362 traditional marker gene analysis. 363 Annotations for reference genomes of homologous protein domains identified by SCOP and 364 other protein-classification schemes, as well as tools for identifying corresponding sequence 365 signatures, are readily available in public databases. An added advantage is that the biochemical 366 function and molecular phenotype of the domains are readily accessible as well, through additional 367 resources including protein data bank (PDB) and InterPro. For complex characters such as protein 368 domains, character homology can be determined with high confidence using sophisticated statistical 369 models (HMMs). Homology of a protein domain implies that the de novo evolution of a genomic 370 locus corresponding to that protein domain is a unique historical event. Therefore, homoplasy 371 due to convergences and parallelisms is highly improbable (68, 69). Although a handful of cases of 372 convergent evolution of 3D structures is known, these instances relate to relatively simple 3D folds 373 coded for by relatively simple sequence repeats (70). 374 However, the vast majority of domains identified by SCOP correspond to polypeptides that are 375 on average 200 residues long with unique sequence profiles (57, 68). Thus, identifying homoplasy 376 in the protein-domain datasets depends largely on estimating reversals, which in this case will 377 be cases of secondary gains/losses; for instance gain-loss-regain events caused by DL-HT or HT. 378 Such secondary gains are more likely to correspond to HT events than to convergent evolution, for 379 reasons specified above. Instances of reversals are minimal, as seen from the strong directional 380 trends detected in the data (Fig. 5B and Fig. 6B).

381 Vertical and horizontal classification 382 For decades, biologists have been faced with a choice between so-called horizontal (Linnean) and 383 vertical (Darwinian) classification of biodiversity (71). The similarity of both schools of systematics 384 concerns the identification of “signatures” or sets of characteristic features that codify evolutionary 385 relationships (54, 63, 71). But the former emphasizes the unity of contemporary groups, i.e. those 386 at a similar evolutionary state, and therefore separates ancestors from descendants, while the latter 387 emphasizes the unity of the ancestors and separates descendants that diverge from a common 388 ancestry (71). Vertical classification is more consistent with the concept of lineal descent, and 389 is the predominant paradigm for which the operational methodology and the algorithmic logic 390 were laid out as the principles of phylogenetic systematics (63, 72). Accordingly, determining the 391 ancestor-descendant polarity, starting from the universal common ancestor (UCA) at the root of the 392 Tree of Life, is crucial to accurately reconstructing the path of evolutionary descent. 393 The classical rooting of the (rRNA) ToL based on the EF-Tu—EF-G paralogous pair (73, 74) is known 394 to be error-prone and highly ambiguous, due to LBA artifacts (14, 75). Remarkably, sequences 395 corresponding to only one of the two conserved domains common to EF-Tu and EF-G ( 200 residues 396 in the P-loop-containing NTP hydrolase domain (Fig. 5A)) can be aligned with confidence (14). 397 Implementing better-fitting substitution models results in two alternative rootings (R1 and R4 in 398 Fig. 5), which relate to distinct, irreconcilable scenarios (14) similar to scenarios in Fig 4C and 4F. 399 Moreover, the EF-Tu—EF-G paralogous pair is only 2 of 57 known paralogs of the translational 400 GTPase protein superfamily (76). Thus the assumption that EF-Tu—EF-G duplication is a unique 401 event, which is essential for the paralogous outgroup-rooting method, is untenable. 402 In the absence of prior knowledge of outgroups or of fossils, rooting the Tree of Life is arguably 403 one of the most difficult phylogenetic problems. Incorrect rooting may lead to profoundly mislead- 404 ing conclusions about evolutionary scenarios and taxonomic affinities, and it appears to be common 405 in phylogenetic studies (77). Perhaps worse yet seems to be the preponderance of subjective a 406 posteriori rooting based on untested preconceptions (e.g. (78, 79)) and scenario-driven erection 407 of taxonomic ranks (e.g. (1, 30)) (80). The conventional practice of a posteriori rooting, wherein an

14 of 21 408 unrooted tree is converted into a rooted tree by adding an ad hoc root, encourages a subjective 409 interpretation of the ToL. For example, the so-called bacterial rooting of the ToL (root R3; Fig. 4) is 410 the preferred rooting hypothesis to interpret the ToL even though that rooting is not well supported 411 (14).

412 Untangling data bias, model bias and investigator bias (prior beliefs) 413 Phylogenies, and hence the taxonomies and evolutionary scenarios they support, are falsifiable 414 hypotheses. Statistical hypothesis testing is now an integral part of phylogenetic inference, to 415 quantify the empirical evidence in support of the various plausible evolutionary scenarios. However, 416 common statistical models implemented for phylogenomic analyses are limited to modeling varia- 417 tion in patterns of point mutations, particularly substitution mutations. These statistical models are 418 intimately linked to basic concepts of molecular evolution, such as the universal molecular clock 419 (3), the universal chronometer (78), paralogous outgroup rooting (81), etc., which are gene-centric 420 concepts that were developed to study the gene, during the age of the gene. Moreover, these 421 idealized notions originated from the analyses of relatively small single-gene datasets. 422 Conventional phylogenomics of multi-locus datasets is a direct extension of the concepts and 423 methods developed for single-locus datasets, which rely exclusively on substitution mutations 424 (50). In contrast, the fundamental concepts of phylogenetic theory: homology, synapomorphy, 425 homoplasy, character polarity, etc., even if idealized, are more generally applicable. And, apparently 426 they are better suited for unique and complex genomic characters rather than for redundant, 427 elementary sequence characters, with regards to determining both qualitative as well as statistical 428 consistency of the data and the underlying assumptions. 429 Phylogenetic theory that was developed to trace the evolutionary history of organismal species, 430 as well as related methods of discrete character analysis for classifying organismal families (63, 82), 431 was adopted, although not entirely, to determine the evolution and classification of gene families (1, 432 3). The discovery and initial description of the Archaea was based on the comparative analysis of a 433 single-gene (rRNA) family. However, in spite of the large number of characters that can be analyzed, 434 neither the rRNA genes nor multi-gene concatenations of core-genes have proved to be efficient 435 phylogenetic markers to reliably resolve the evolutionary history and phylogenetic affinities of the 436 Archaea (83, 84). 437 Uncertainties and errors in phylogenetic inference are primarily errors in adequately distinguish- 438 ing homologous similarities from homoplastic similarities (34, 50, 85). Homologies, synapomorphies 439 and homoplasies are qualitative inferences, yet are inherently statistical (probabilistic). The prob- 440 abilistic framework (maximum likelihood and Bayesian methods) has proven to be powerful for 441 quantifying uncertainties and testing alternative hypotheses. Log odds ratios, such as LLR and 442 LBF, are measures of how one changes belief in a hypothesis in light of new evidence (86). Accord- 443 ingly, directional evolution models are more optimal explanations of the observed distribution of 444 genomic-characters, and such directional trends overwhelmingly support the monophyly of the 445 Archaea, as well as the sisterhood of the Archaea and the Bacteria, i.e. monophyly of Akarya (Fig 6). 446 Data quality is at least as important as the evolution models that are posited to explain the 447 data. Although sophisticated statistical tests for evaluating tree robustness, and for selecting 448 character-evolution models, are becoming a standard feature of phylogenetic software (e.g. IQ-tree, 449 MrBayes, Phylobayes), tests for character evaluation are not common. Routines for collecting and 450 curating data upstream of phylogenetic analyses are rather eclectic. Besides, it is an open question 451 as to whether qualitatively different datasets (as in Fig.1 and Fig.2) can be compared effectively. 452 Nevertheless, employing DDNs and other tools of exploratory data analysis could be useful to 453 identify conflicts that arise due to data collection and/or curation errors (23, 24).

454 Conclusions 455 The Tree of Life is primarily a phylogenetic classiﬁcation that is invaluable to organize and to 456 describe the evolution of biodiversity, explicated through evolutionary scenarios. Phylogenies are

15 of 21 457 hypotheses that mostly relate to extinct ancestors, while taxonomies are hypotheses that largely 458 relate to extant species. Extant species contain distinct combinatorial mosaics of ancestral features 459 (plesiomorphies) and evolutionary novelties (apomorphies). It is remarkable that the uniqueness of 460 the Archaea was identified by the comparative analyses of oligonucleotide signatures in a single 461 gene dataset (1). However the same is not true of the phylogenetic classification of the Archaea, 462 based on marker-genes and reversible evolution models that rely exclusively on point mutations, 463 specifically substitution mutations, which may not be ideal phylogenetic markers (59). 464 The Three-domains of Life hypothesis (26), which was initially based on the interpretation of an 465 unrooted rRNA tree (of life) (1), was put forward largely to emphasize the uniqueness of the Archaea, 466 ascribed to an exclusive lineal descent. Although many lines of evidence, molecular or otherwise, 467 support the uniqueness of the Archaea, phylogenetic analysis of genomic signatures does not 468 support the presumed primitive state of Archaea or Bacteria, and the common belief that Archaea 469 and Bacteria are ancestors of Eukarya (1, 11, 39, 87). Models of evolution of genomic features 470 support a Two-domains (or rather two empires) of Life hypothesis (9), as well as the independent 471 origins and parallel descent of eukaryote and akaryote species (10, 14, 88, 89).

472 Data and methods 473 Data collection and curation Marker domains datasets 474

475 Character matrices of homologous protein-domains, coded as binary-state characters were assem- 476 bled from genome annotations of SCOP-domains available through the SUPERFAMILY HMM library 477 and genome assignments server; v. 1.75 (http://supfam.org/SUPERFAMILY/) (57, 90). 478 (i) 141-species dataset was obtained from a previous study (29) 479 (ii) The 141-species dataset was updated with representatives of novel species described recently, 480 largely with archaeal species from TACK group (30), DPANN group (5) and Asgard group including 481 the Lokiarchaeota (20). In addition, species sampling was enhanced with representatives from 482 the candidate phyla (unclassiﬁed) described for bacterial species and with unicellular species of 483 eukaryotes, to a total of 222 species. The complete list of the species with their respective Taxonomy 484 IDs is available in SI Table 1. 485 When genome annotations were unavailable from SUPERFAMILY database, curated reference 486 proteomes were obtained from the universal protein resource (http://www.uniprot.org/proteomes/). 487 SCOP-domains were annotated using the HMM library and genome annotation tools and routines 488 recommended by the SUPERFAMILY resource. Marker genes datasets 489

490 Marker gene datasets from previous studies were obtained as follows, (i) 29 core-genes align- 491 ment(17) and (ii) SSU rRNA alignment and 48 core-genes alignments (20).

492 Exploratory data analysis 493 DDNs were constructed with SplitsTree v. 4.14. Split networks were computed using the Neigh- 494 borNet method from the observed P-distances of the taxa for both nucleotide- and amino acid- 495 characters. Split networks of the protein-domain characterss were computed from Hamming 496 distance, which is identical to the P-distance. The networks were drawn with the equal angle 497 algorithm.

498 Phylogenetic analyses 499 Concatenated gene tree inference: Extensive analyses of the concatenated core-genes datasets 500 are reported in the original studies (17, 20). Analysis here was restricted to the 29 core-genes 501 dataset due its relatively small taxon sampling (44 species) compared to the 48 core-genes dataset 502 (96 species) since there is little diﬀerence in data quality, but the computational time/resources

16 of 21 503 required is significantly lesser. Moreover, the general conclusions based on these datasets are 504 consistent despite a smaller taxon sampling, particularly of archaeal species (26 as opposed to 64 505 in the larger sampling). 506 Best-fitting amino acid substitution models were chosen using Smart Model Selection (SMS) 507 (91) compatible with PhyML tree inference methods (92). Trees were estimated with a rate- 508 homogeneous LG model as well as rate-heterogeneous versions of the LG model. Site-specific rate 509 variation was approximated using the gamma distribution with 4, 8 and 12 rate categories, LG+G4, 510 LG+G8 and LG+G12, respectively. More complex models (SI Table 2) that account for invariable sites 511 (LG+GX+I) and/or models that compute alignment-specific state frequencies (LG+GX+F) were also 512 used, but the trees inferred were identical to trees estimated from LG+GX models, and therefore not 513 reported here. Log likelihoods ratio (LLR) was calculated as the difference in the raw log likelihoods 514 for each model. 515 Genome tree inference: The Mk model (32) is the most widely implemented model for phyloge- 516 netic inference in the probabilistic framework (maximum likelihood (ML) and Bayesian methods) 517 applicable to complex features coded as binary characters. However, only the reversible model is 518 implemented in ML methods at present. Both reversible and directional evolution models as well as 519 model selection routines implemented in MrBayes 3.2 (42, 93) were used. The Metropolis-coupled 520 MCMC algorithm was used with two chains, sampling every 500th generation. The first half of 521 the generations was discarded as burn-in. MCMC sampling was run until convergence, unless 522 mentioned otherwise. Convergence was assessed through the average standard deviation of 523 split frequencies (ASDSF, less than 0.01) for tree topology and the potential scale reduction factor 524 (PSRF, equal to 1.00) for scalar parameters, unless mentioned otherwise. Bayes factors for model 525 comparison were calculated using the harmonic mean estimator in MrBayes. The log Bayes factor 526 (LBF) was calculated as the difference in the log likelihoods for each model. 527 Convergence between independent runs was generally slower for directional models compared 528 to the reversible models. When convergence was extremely slow (requiring more than 100 million 529 generations) topology constraints corresponding to the clusters derived in the unrooted trees (Fig. 530 3E) were applied to improve convergence rates. In general these clusters/constraints corresponded 531 to named taxonomic groups e.g. Fungi, Metazoa, Crenarchaeota, etc. Convergence assessment 532 between independent runs was relaxed for three specific cases that did not converge at the time of 533 submission: the unrooted tree with Mk-uniform-rates model (ASDSF 0.05; PSRF 1.03), rooted trees 534 corresponding to root-R2 (ASDSF 0.5; PSRF 1.04) and root-R3 (ASDSF 0.029; PSRF 1.03). In the three 535 cases specified, the difference in bipartitions is in the shallow parts (minor branches) of the tree. 536 For assessing well supported major branches of the tree, ASDSF values between 0.01 and 0.05 may 537 be adequate, as recommended by the authors (94).

538 Funding 539 This research received no speciﬁc grant from any funding agency in the public, commercial, or 540 not-for-proﬁt sectors. Work by this author was partially supported by The Swedish Research Council 541 (to Måns Ehrenberg) and the Knut and Alice Wallenberg Foundation, RiboCORE (to Måns Ehrenberg 542 and Dan Andersson).

543 Acknowledgements 544 I am grateful to Charles (Chuck) Kurland and Måns Ehrenberg for support and encouragement. 545 I thank Chuck Kurland and Siv Andersson for the discussions in general; Chuck for the many 546 stimulating debates and Siv for inspiring the article title, in part; Seraina Klopfstein for providing 547 the algorithms for implementing the directional model in MrBayes and for helpful suggestions and 548 Erling Wikman for help with computing equipment.

17 of 21 549 References 550 1. Woese CR, Fox GE. Phylogenetic structure of the prokaryotic domain: the primary kingdoms. 551 Proceedings of the National Academy of Sciences. 1977;74(11):5088-90. 552 2. Woese CR. The Archaeal Concept and the World it Lives in: A Retrospective. Photosynthesis 553 Research. 2004;80(1):361-72. 554 3. Zuckerkandl E, Pauling L. Molecules as documents of evolutionary history. Journal of theoreti- 555 cal biology. 1965;8(2):357-66. 556 4. Ragan MA, Bernard G, Chan CX. Molecular phylogenetics before sequences. RNA Biology. 557 2014;11(3):176-85. 558 5. Rinke C, Schwientek P, Sczyrba A, Ivanova NN, Anderson IJ, Cheng J-F, et al. Insights into the 559 phylogeny and coding potential of microbial dark matter. Nature. 2013;499(7459):431-7. 560 6. Sunagawa S, Coelho LP, Chaﬀron S, Kultima JR, Labadie K, Salazar G, et al. Structure and 561 function of the global ocean microbiome. Science. 2015;348(6237):1261359. 562 7. Wu D, Wu M, Halpern A, Rusch DB, Yooseph S, Frazier M, et al. Stalking the Fourth Domain in 563 Metagenomic Data: Searching for, Discovering, and Interpreting Novel, Deep Branches in Marker 564 Gene Phylogenetic Trees. PLOS ONE. 2011;6(3):e18011. 565 8. Boyer M, Madoui M-A, Gimenez G, La Scola B, Raoult D. Phylogenetic and Phyletic Studies 566 of Informational Genes in Genomes Highlight Existence of a 4th Domain of Life Including Giant 567 Viruses. PLOS ONE. 2010;5(12):e15530. 568 9. Mayr E. Two empires or three? Proceedings of the National Academy of Sciences of the United 569 States of America. 1998;95(17):9720-3. 570 10. Harish A, Tunlid A, Kurland CG. Rooted phylogeny of the three superkingdoms. Biochimie. 571 2013;95(8):1593-604. 572 11. Williams TA, Foster PG, Cox CJ, Embley TM. An archaeal origin of eukaryotes supports only 573 two primary domains of life. Nature. 2013;504(7479):231-6. 574 12. Hug LA, Baker BJ, Anantharaman K, Brown CT, Probst A J, Castelle CJ, et al. A new view of the 575 tree of life. Nature Microbiology. 2016;1:16048. 576 13. Parks DH, Rinke C, Chuvochina M, Chaumeil P-A, Woodcroft BJ, Evans PN, et al. Recovery 577 of nearly 8,000 metagenome-assembled genomes substantially expands the tree of life. Nature 578 Microbiology. 2017. 579 14. Gouy R, Baurain D, Philippe H. Rooting the tree of life: the phylogenetic jury is still out. Phil 580 Trans R Soc B. 2015;370(1678):20140329. 581 15. Lake JA. An alternative to archaebacterial dogma. Nature. 1986;319(6055):626-. 582 16. Tourasse NJ, Gouy M. Accounting for evolutionary rate variation among sequence sites con- 583 sistently changes universal phylogenies deduced from rRNA and protein-coding genes. Molecular 584 phylogenetics and evolution. 1999;13(1):159-68. 585 17. Williams TA, Embley TM. Archaeal "dark matter" and the origin of eukaryotes. Genome 586 Biology and Evolution. 2014;6(3):474-81. 587 18. Spang A, Saw JH, Jørgensen SL, Zaremba-Niedzwiedzka K, Martijn J, Lind AE, et al. Complex 588 archaea that bridge the gap between prokaryotes and eukaryotes. Nature. 2015;521(7551):173-9. 589 19. Da Cunha V, Gaia M, Gadelle D, Nasir A, Forterre P. Lokiarchaea are close relatives of Eur- 590 yarchaeota, not bridging the gap between prokaryotes and eukaryotes. PLOS Genetics. 2017;13(6):e1006810. 591 20. Zaremba-Niedzwiedzka K, Caceres EF, Saw JH, Bäckström D, Juzokaite L, Vancaester E, et al. 592 Asgard archaea illuminate the origin of eukaryotic cellular complexity. Nature. 2017;541(7637):353- 593 8. 594 21. Zwickl DJ, Hillis DM. Increased taxon sampling greatly reduces phylogenetic error. Systematic 595 Biology. 2002;51(4):588-98. 596 22. Salichos L, Rokas A. Inferring ancient divergences requires genes with strong phylogenetic 597 signals. Nature. 2013;497(7449):327-31.

18 of 21 598 23. Morrison DA. Using data-display networks for exploratory data analysis in phylogenetic 599 studies. Molecular Biology and Evolution. 2009;27(5):1044-57. 600 24. Huson DH, Bryant D. Application of phylogenetic networks in evolutionary studies. Mol Biol 601 Evol. 2006;23. 602 25. Woese CR. On the evolution of cells. Proceedings of the National Academy of Sciences of the 603 United States of America. 2002;99(13):8742-7. 604 26. Woese CR, Kandler O, Wheelis ML. Towards a natural system of organisms: Proposal for the 605 domains Archaea, Bacteria, and Eucarya. Proceedings of the National Academy of Sciences of the 606 United States of America. 1990;87(12):4576-9. 607 27. Garrett RA. Molecular evolution: The uniqueness of Archaebacteria. Nature. 1985;318:233-5. 608 28. Valentine DL. Adaptations to energy stress dictate the ecology and evolution of the Archaea. 609 Nature Reviews Microbiology. 2007;5(4):316-23. 610 29. Harish A, Kurland CG. Akaryotes and Eukaryotes are independent descendants of a universal 611 common ancestor. Biochimie. 2017;138:168-83. 612 30. Guy L, Ettema TJG. The archaeal TACK superphylum and the origin of eukaryotes. Trends in 613 microbiology. 2011;19(12):580-7. 614 31. Le SQ, Gascuel O. An Improved General Amino Acid Replacement Matrix. Molecular Biology 615 and Evolution. 2008;25(7):1307-20. 616 32. Lewis PO. A Likelihood Approach to Estimating Phylogeny from Discrete Morphological 617 Character Data. Systematic Biology. 2001;50(6):913-25. 618 33. Wright AM, Hillis DM. Bayesian Analysis Using a Simple Likelihood Model Outperforms Parsi- 619 mony for Estimation of Phylogeny from Discrete Morphological Data. PLOS ONE. 2014;9(10):e109210. 620 34. Morrison DA. Phylogenetic Analyses of Parasites in the New Millennium. Advances in 621 Parasitology2006. p. 1-124. 622 35. Wiley EO, Lieberman BS. Phylogenetics: theory and practice of phylogenetic systematics: 623 John Wiley & Sons; 2011. 624 36. Whittaker RH. New concepts of kingdoms of organisms. Science. 1969;163(3863):150-60. 625 37. Nasir A, Kim K, Caetano-Anolles G. Giant viruses coexisted with the cellular ancestors and 626 represent a distinct supergroup along with superkingdoms Archaea, Bacteria and Eukarya. BMC 627 Evolutionary Biology. 2012;12(1):156. 628 38. Stanier RY, Niel Cv. The concept of a bacterium. Archives of Microbiology. 1962;42(1):17-35. 629 39. Sagan L. On the origin of mitosing cells. Journal of theoretical biology. 1967;14(3):225-75. 630 40. Yang Z, Roberts D. On the use of nucleic acid sequences to infer early branchings in the tree 631 of life. Molecular Biology and Evolution. 1995;12(3):451-8. 632 41. Huelsenbeck JP, Bollback JP, Levine AM. Inferring the root of a phylogenetic tree. Systematic 633 biology. 2002;51(1):32-43. 634 42. Klopfstein S, Vilhelmsen L, Ronquist F. A Nonstationary Markov Model Detects Directional 635 Evolution in Hymenopteran Morphology. Systematic Biology. 2015;64(6):1089-103. 636 43. Criscuolo A, Gribaldo S. BMGE (Block Mapping and Gathering with Entropy): a new soft- 637 ware for selection of phylogenetic informative regions from multiple sequence alignments. BMC 638 Evolutionary Biology. 2010;10(1):210. 639 44. Light S, Sagit R, Sachenkova O, Ekman D, Elofsson A. Protein expansion is primarily due to 640 indels in intrinsically disordered regions. Molecular Biology and Evolution. 2013;30(12):2645-53. 641 45. Murzin AG, Brenner SE, Hubbard T, Chothia C. SCOP: A structural classiﬁcation of pro- 642 teins database for the investigation of sequences and structures. Journal of Molecular Biology. 643 1995;247(4):536-40. 644 46. Woese CR, Fox GE, Zablen L, Uchida T, Bonen L, Pechman K, et al. Conservation of primary 645 structure in 16S ribosomal RNA. Nature. 1975;254(5495):83-6. 646 47. Rokas A, Carroll SB. Frequent and widespread parallel evolution of protein sequences. 647 Molecular Biology and Evolution. 2008;25(9):1943-53.

19 of 21 648 48. Parker J, Tsagkogeorga G, Cotton JA, Liu Y, Provero P, Stupka E, et al. Genome-wide signatures 649 of convergent evolution in echolocating mammals. Nature. 2013;502(7470):228-31. 650 49. Rokas A, Carroll SB. Bushes in the tree of life. PLoS Biology. 2006;4(11):1899-904. 651 50. Philippe H, Roure B. Difficult phylogenetic questions: more data, maybe; better methods, 652 certainly. BMC Biology. 2011;9(1):1-4. 653 51. Shen X-X, Hittinger CT, Rokas A. Contentious relationships in phylogenomic studies can be 654 driven by a handful of genes. Nature ecology & evolution. 2017;1(5):0126. 655 52. Springer MS, Gatesy J. On the importance of homology in the age of phylogenomics. 656 Systematics and Biodiversity. 2017:1-19. 657 53. Govindarajan S, Recabarren R, Goldstein RA. Estimating the total number of protein folds. 658 Proteins: Structure, Function and Genetics. 1999;35(4):408-14. 659 54. Graham DE, Overbeek R, Olsen GJ, Woese CR. An archaeal genomic signature. Proceedings 660 of the National Academy of Sciences. 2000;97(7):3304-8. 661 55. Park J, Karplus K, Barrett C, Hughey R, Haussler D, Hubbard T, et al. Sequence comparisons 662 using multiple sequences detect three times as many remote homologues as pairwise methods. 663 Journal of Molecular Biology. 1998;284(4):1201-10. 664 56. Eddy SR. Accelerated profile HMM searches. PLoS Computational Biology. 2011;7(10). 665 57. Gough J, Karplus K, Hughey R, Chothia C. Assignment of homology to genome sequences 666 using a library of hidden Markov models that represent all proteins of known structure. Journal of 667 Molecular Biology. 2001;313(4):903-19. 668 58. Fang H, Oates ME, Pethica RB, Greenwood JM, Sardar A J, Rackham OJL, et al. A daily-updated 669 tree of (sequenced) life as a reference for genome research. Scientific Reports. 2013;3. 670 59. Rokas A, Holland PWH. Rare genomic changes as a tool for phylogenetics. Trends in Ecology 671 and Evolution. 2000;15(11):454-9. 672 60. Castoe TA, de Koning A J, Pollock DD. Adaptive molecular convergence: Molecular evolution 673 versus molecular phylogenetics. Communicative and Integrative Biology. 2010;3(1):12-7. 674 61. Liu Y, Cotton JA, Shen B, Han X, Rossiter SJ, Zhang S. Convergent sequence evolution between 675 echolocating bats and dolphins. Current Biology. 2010;20(2):R53-R4. 676 62. Foote AD, Liu Y, Thomas GWC, Vinar T, Alfoldi J, Deng J, et al. Convergent evolution of the 677 genomes of marine mammals. Nat Genet. 2015;advance online publication. 678 63. Hennig W. Phylogenetic systematics. Annual review of entomology. 1965;10(1):97-116. 679 64. Padian K, Lindberg DR, Polly PD. Cladistics and the fossil record: the uses of history. Annual 680 Review of Earth and Planetary Sciences. 1994;22:63-91. 681 65. Lienau EK, DeSalle R. Is the microbial tree of life verificationist? Cladistics. 2010;26(2):195-201. 682 66. Sanderson MJ. Where have all the clades gone? A systematist’s take in Inferring Phylogenies. 683 Evolution. 2005;59(9):2056-8. 684 67. Wheeler Q, Assis L, Rieppel O. Phylogenetics: Heed the father of cladistics. Nature. 685 2013;496(7445):295-6. 686 68. Pethica RB, Levitt M, Gough J. Evolutionarily consistent families in SCOP: Sequence, structure 687 and function. BMC Structural Biology. 2012;12. 688 69. Mackin KA, Roy RA, Theobald DL. An empirical test of convergent evolution in rhodopsins. 689 Molecular Biology and Evolution. 2014;31(1):85-95. 690 70. Mistry J, Finn RD, Eddy SR, Bateman A, Punta M. Challenges in homology search: HMMER3 691 and convergent evolution of coiled-coil regions. Nucleic acids research. 2013;41(12):e121-e. 692 71. Simpson GG. The Principles of Classification and a Classification of Mammals. Bull Amer 693 Museum Nat History. 1945;85:xvi+350. 694 72. Felsenstein J. Inferring phylogenies. Sunderland, MA: Sinauer Associates; 2004. 695 73. Iwabe N, Kuma K, Hasegawa M, Osawa S, Miyata T. Evolutionary relationship of archaebacte- 696 ria, eubacteria, and eukaryotes inferred from phylogenetic trees of duplicated genes. Proceedings 697 of the National Academy of Sciences. 1989;86(23):9355-9.

20 of 21 698 74. Baldauf SL, Palmer JD, Doolittle WF. The root of the universal tree and the origin of eukaryotes 699 based on elongation factor phylogeny. Proceedings of the National Academy of Sciences of the 700 United States of America. 1996;93(15):7749-54. 701 75. Forterre P, Philippe H. Where is the root of the universal tree of life? BioEssays. 1999;21(10):871- 702 9. 703 76. Atkinson GC. The evolutionary and functional diversity of classical and lesser-known cyto- 704 plasmic and organellar translational GTPases across the tree of life. BMC Genomics. 2015;16(1):78. 705 77. Graham SW, Olmstead RG, Barrett SCH. Rooting Phylogenetic Trees with Distant Outgroups: 706 A Case Study from the Commelinoid Monocots. Molecular Biology and Evolution. 2002;19(10):1769- 707 81. 708 78. Woese CR. Bacterial evolution. Microbiological reviews. 1987;51(2):221. 709 79. Nasir A, Caetano-Anollés G. A phylogenomic data-driven exploration of viral origins and 710 evolution. Science Advances. 2015;1(8). 711 80. Gribaldo S, Brochier-Armanet C. Time for order in microbial systematics. Trends in microbi- 712 ology. 2012;20(5):209-10. 713 81. Schwartz R, Dayhoff M. Origins of prokaryotes, eukaryotes, mitochondria, and chloroplasts. 714 Science. 1978;199(4327):395-403. 715 82. Darwin C. On the Origin of Species by Means of Natural Selection, or the Preservation of 716 Favoured Races in the Struggle for Life. London: John Murray; 1859. 717 83. Gribaldo S, Poole AM, Daubin V, Forterre P, Brochier-Armanet C. The origin of eukaryotes 718 and their relationship with the Archaea: are we at a phylogenomic impasse? Nat Rev Micro. 719 2010;8(10):743-52. 720 84. Gupta RS. Impact of genomics on the understanding of microbial evolution and classification: 721 the importance of Darwin’s views on classification. FEMS microbiology reviews. 2016;40(4):520-53. 722 85. Avise JC, Robinson TJ. Hemiplasy: a new term in the lexicon of phylogenetics. Systematic 723 Biology. 2008;57(3):503-7. 724 86. Huelsenbeck JP, Larget B, Alfaro ME. Bayesian Phylogenetic Model Selection Using Reversible 725 Jump Markov Chain Monte Carlo. Molecular Biology and Evolution. 2004;21(6):1123-33. 726 87. Woese CR. Interpreting the universal phylogenetic tree. Proceedings of the National Academy 727 of Sciences. 2000;97(15):8392-6. 728 88. Brinkmann H, Philippe H. Archaea sister group of Bacteria? Indications from tree reconstruc- 729 tion artifacts in ancient phylogenies. Molecular biology and evolution. 1999;16(6):817-25. 730 89. Harish A, Kurland CG. Mitochondria are not captive bacteria. Journal of Theoretical Biology. 731 2017;434:88-98. 732 90. Oates ME, Stahlhacke J, Vavoulis DV, Smithers B, Rackham OJL, Sardar AJ, et al. The SUPER- 733 FAMILY 1.75 database in 2014: A doubling of data. Nucleic Acids Research. 2015;43(D1):D227-D33. 734 91. Lefort V, Longueville J-E, Gascuel O. SMS: Smart Model Selection in PhyML. Molecular Biology 735 and Evolution. 2017:msx149. 736 92. Guindon S, Dufayard J-F, Lefort V, Anisimova M, Hordijk W, Gascuel O. New algorithms and 737 methods to estimate maximum-likelihood phylogenies: assessing the performance of PhyML 3.0. 738 Systematic biology. 2010;59(3):307-21. 739 93. Ronquist F, Teslenko M, van der Mark P, Ayres DL, Darling A, Höhna S, et al. MrBayes 3.2: 740 Efficient Bayesian Phylogenetic Inference and Model Choice Across a Large Model Space. Systematic 741 Biology. 2012;61(3):539-42. 742 94. Ronquist F, Huelsenbeck J, Teslenko M. MrBayes version 3.2 manual: tutorials and model 743 summaries. Available with the software distribution at mrbayessourceforgenet/mb32_manualpdf. 744 2011.